Exploring the Dynamics of Data Transmission in 5G Networks: A Conceptual Analysis

Nikita Smirnov, Sven TomfordearXiv 2024 · DOI · arXiv

Abstract

Maintaining high-quality, real-time data streams is critical for the safe operation of autonomous maritime vessels. This paper investigates the challenges of transmitting high-bandwidth video and LiDAR data from remote-controlled units to shore-based control centers over 5G networks. We present a novel framework, Gymir5G, which integrates a custom-built 5G simulation environment with a Deep Reinforcement Learning (DRL) approach for adaptive congestion control. Our findings indicate that AI-driven models can significantly outperform traditional congestion control algorithms, such as Google Congestion Control (GCC), by proactively managing bitrates in fluctuating maritime network conditions.

Introduction: The Challenge of Maritime Situational Awareness

Remote-controlled units (RCUs), exemplified by the "Wavelab" ferry operating in Kiel, Germany, necessitate robust "situational awareness" to ensure safe navigation and docking. This requirement translates to the continuous streaming of multiple 360-degree high-definition video feeds and dense LiDAR point clouds to human operators or autonomous systems on shore.

The "Wavelab" autonomous ferry used for real-world measurements in Kiel, Germany.

The primary technical hurdle is the unpredictable nature of 5G networks in maritime environments. Factors such as base station handovers, signal degradation over open water, and network congestion create a volatile environment for data transmission. Furthermore, traditional streaming protocols often fail to meet the sub-second latency requirements of real-time docking maneuvers. TCP-based protocols prioritize reliability at the cost of speed, while modern alternatives like WebRTC utilize congestion control algorithms that are frequently too conservative, leading to underutilized bandwidth and potential stalling.

Methodology: The Gymir5G Framework

To address these challenges, we developed a multi-layered software stack designed to analyze and optimize the data pipeline. Central to this approach is Gymir5G, an OMNeT++-based simulator tailored for 5G maritime scenarios.

The Gymir5G simulation architecture, bridging the C++ simulation world with the Python DRL environment.

The architecture bridges two distinct domains: the C++ SIM world, which handles the network simulation (Simu5G) and control station logic, and the Python DRL world, where the reinforcement learning agent resides. A dedicated "Communicator" module facilitates the exchange of state information (e.g., receive rate, RTT, packet loss) and actions (e.g., bitrate, resolution, compression) between the two environments.

Key components of our methodology include:

  • WebRTC Optimization: We identified WebRTC as the optimal protocol for achieving sub-300ms latency. We enhanced its performance using Negative Acknowledgment (NACK) for targeted retransmissions and Forward Error Correction (FEC) to provide redundancy against packet loss.
  • LiDAR-as-Video: To integrate high-density LiDAR data into the standard WebRTC pipeline, we rendered point clouds into video frames at the source. This allows LiDAR data to be treated as a standard media stream, leveraging existing hardware decoders.
  • AI-Enhanced Congestion Control: We implemented a Soft Actor-Critic (SAC) DRL model. Unlike the standard Google Congestion Control (GCC), which relies on trendline filters, the SAC agent learns from simulated experience to take proactive risks, maintaining higher throughput even in the face of minor network fluctuations.

Key Insights and Experimental Results

Our analysis yielded several critical insights into the dynamics of 5G maritime data transmission:

Protocol and Codec Performance

Despite the emergence of newer codecs like H.265 and AV1, hardware-accelerated H.264 remains the most viable choice for real-time 5G streaming. Its stability, sub-1ms encoding latency, and broad browser support outweigh the compression advantages of more modern alternatives in current hardware ecosystems.

DRL vs. Traditional Congestion Control

The DRL approach demonstrated a clear advantage over the standard GCC. In "moderate" congestion scenarios, GCC often struggled to exceed 3 Mbps due to its conservative response to packet loss. In contrast, the DRL agent learned to maintain higher bitrates by accepting minimal packet loss, resulting in a more consistent and higher-quality stream for the operator.

Network Infrastructure: 5G SA vs. NSA

Our simulations and real-world measurements from the "Wavelab" ferry highlighted the superiority of 5G Standalone (SA) architecture. In maritime uplink scenarios, 5G SA provided approximately three times the performance of Non-Standalone (NSA) configurations. Additionally, while Wi-Fi is a useful supplement, its effective range was limited to roughly 80 meters from the coast, making it suitable only for final docking phases.

Conclusion and Future Directions

This study demonstrates that integrating AI-driven congestion control with optimized 5G protocols can meet the stringent requirements of autonomous maritime navigation. The total pipeline latency was successfully maintained under 300 milliseconds, a critical threshold for remote control.

However, challenges remain. The performance of DRL models is highly sensitive to the design of the reward function, and further research is required to ensure these models generalize across different geographic locations and extreme weather conditions. Future work will also focus on overcoming the browser integration and stability issues of H.265 and AV1 to support future 8K streaming requirements.

Summary updated

BibTeX
@misc{smirnov2024exploring,
  author = {Nikita Smirnov and Sven Tomforde},
  title = {{Exploring the Dynamics of Data Transmission in 5G Networks: A Conceptual Analysis}},
  year = {2024},
  eprint = {2404.16508},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2404.16508},
}
Figure